US5548697A - Non-linear color corrector having a neural network and using fuzzy membership values to correct color and a method thereof - Google Patents

Non-linear color corrector having a neural network and using fuzzy membership values to correct color and a method thereof Download PDF

Info

Publication number
US5548697A
US5548697A US08/367,798 US36779894A US5548697A US 5548697 A US5548697 A US 5548697A US 36779894 A US36779894 A US 36779894A US 5548697 A US5548697 A US 5548697A
Authority
US
United States
Prior art keywords
pixel
pixels
color
image
neural network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
US08/367,798
Other languages
English (en)
Inventor
Anthony E. Zortea
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Panasonic Corp of North America
Original Assignee
Panasonic Technologies Inc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Panasonic Technologies Inc filed Critical Panasonic Technologies Inc
Priority to US08/367,798 priority Critical patent/US5548697A/en
Assigned to PANASONIC TECHNOLOGIES, INC. reassignment PANASONIC TECHNOLOGIES, INC. ASSIGNMENT OF ASSIGNORS INTEREST (SEE DOCUMENT FOR DETAILS). Assignors: ZORTEA, ANTHONY E.
Priority to EP95120188A priority patent/EP0720390B1/en
Priority to DE69517401T priority patent/DE69517401T2/de
Priority to JP7341926A priority patent/JPH08256349A/ja
Application granted granted Critical
Publication of US5548697A publication Critical patent/US5548697A/en
Assigned to MATSUSHITA ELECTRIC CORPORATION OF AMERICA reassignment MATSUSHITA ELECTRIC CORPORATION OF AMERICA MERGER (SEE DOCUMENT FOR DETAILS). Assignors: PANASONIC TECHNOLOGIES, INC.
Anticipated expiration legal-status Critical
Expired - Fee Related legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N9/00—Details of colour television systems
    • H04N9/64—Circuits for processing colour signals
    • H04N9/646—Circuits for processing colour signals for image enhancement, e.g. vertical detail restoration, cross-colour elimination, contour correction, chrominance trapping filters

Definitions

  • the present invention relates to a non-linear color corrector and more particularly, to a non-linear color corrector which includes a neural network for defining the class of colors to be changed.
  • Non-linear processes can occur at any one of these steps to corrupt the color fidelity of the image.
  • a video camera is made up of three separate channels, usually identified as the red, green, and blue primaries.
  • the required spectral sensitivity of the three camera primaries can be derived using the chromaticities, a physical measurement related to spectral content, of the display device's three primaries.
  • the camera primaries are related to the spectral sensitivity of the prism that separates the light into three channels. Color reproduction fidelity suffers to the extent the camera primaries do not match the three primaries of the display.
  • Non-linear color correction involves several steps including: (1) defining a class of colors that need to be changed (hereinafter referred to as the "change-class"), (2) defining how a pixel belonging to that class is to be changed, (3) examining each pixel and determining the extent to which it belongs to the change-class, and (4) changing each pixel according to its membership in the change-class.
  • change-class a class of colors that need to be changed
  • the change-class can be defined by user examples. For example, the user can examine the image and generate a list of pixels that should undergo significant change, should undergo little change, or should not be changed. In addition, the user can also control how the pixels should be changed. This is a simple, model-free method of acquiring the necessary information from the user.
  • FIG. 1 illustrates the difficulties which may be encountered in defining the change-class using a model-free class membership definition.
  • FIG. 1 is a graph showing a mapping of color components in a two-dimensional color image into two color components.
  • the user's change-class is not a simple shape like an ellipse or a rectangle. As a result, it is difficult to model the shape of this class.
  • the difficulty increases in three dimensions (i.e., when a third color component is used in the mapping operation).
  • the pixels in the change-class are changed once the membership in the change-class is determined.
  • the pixels can be changed using both modelled and model-free methods.
  • FIG. 2 illustrates the modelled method for changing a pixel.
  • the pixels having color combinations near pixel p1, for example pixel p3, are changed towards pixel p2 to either, for example, pixels pa or pb, having color combinations near pixel p2.
  • a color corrector examines pixel p3, determines if it belongs to the change-class, and provides a membership value c designating the pixels membership in the change-class.
  • the pixel p3 can be changed towards p2 to pixel pa by an amount proportional to the membership value c in the direction of vector A according to equation (1) below.
  • pixel p3 can be changed in the same direction as the vector V to color combination pb according to equation (2) below.
  • Equations (1) and (2) essentially move a pixel along a vector by an amount proportional to its membership value c in the pixel change-class. Equation (1) moves the pixel towards the "target" pixel and equation (2) moves the pixel in the same direction as the original vector. Equation (2) maintains color resolution near the colors that will be changed by remapping pixels in the same direction. Equation (1) allows a relatively larger color-volume of pixels to be mapped to a smaller color-volume, thus, decreasing color resolution.
  • an apparatus and method are needed to define the change-class and to remap the pixels in the change-class.
  • the present invention provides a color corrector for changing pixels in an image.
  • the color corrector has a neural fuzzy classifier which generates a membership value that defines the degree of membership of the pixel in a group of pixels to be transformed.
  • a pixel color changer is also provided which transforms the pixel according to its membership in the group of pixels to be changed.
  • the present invention also relates to a pixel group classifier which identifies groups of pixels to train the neural fuzzy classifier to generate the membership value.
  • FIG. 1 is a graph of pixels in an image with respect to two different color bands.
  • FIG. 2 is a graph of a two-dimensional color plane illustrating the prior art modelled method for changing a pixel.
  • FIG. 3 is a block diagram of the color corrector according to an exemplary embodiment of the present invention.
  • FIG. 4a is a block diagram illustrating a conventional feed-forward neural network.
  • FIG. 4b is a block diagram illustrating a neural network in the pixel fuzzy classifier 410 shown in FIG. 3 according to an exemplary embodiment of the present invention.
  • FIG. 5 is a graph of an error surface with only one weight.
  • FIG. 6 is a flow chart diagram of a training method for training the fuzzy neural classifier shown in FIG. 3.
  • FIG. 6a is a graph illustrating training step 530 shown in FIG. 6.
  • FIG. 7 is a graph illustrating the use of the training method in FIG. 6.
  • FIG. 8 is a 3-dimensional plot of the location of an image in an Red-Green-Blue (RGB) cube.
  • RGB Red-Green-Blue
  • FIG. 9 is a two-dimensional graph of the change-class examples and the do-nothing class examples of the imager in FIG. 8 projected onto an Red-Green (RG) plane.
  • RG Red-Green
  • FIG. 10 is another two-dimensional graph of the change-class examples and additional do-nothing class examples of the image in FIG. 8 projected onto an RG plane.
  • FIG. 11 is a block diagram of a Kohonen network in the pixel group classifier 400 shown in FIG. 3.
  • FIG. 12 is two-dimensional graph of change-class examples and do-nothing class examples generated by a Kohonen network of the image in FIG. 8 projected onto an RG plane.
  • FIG. 13 is a graph showing the projection of the location of each pixel in an image onto a two-dimensional color plane.
  • FIG. 14 is a plot showing the projection of the location of each pixel in the image shown in FIG. 13 onto a two-dimensional color plane where the image has been color corrected using a color corrector according to the exemplary embodiment of the present invention.
  • FIG. 15 is a block diagram of a color correction system according to an exemplary embodiment of the present invention.
  • the exemplary embodiment of the present invention includes a pixel fuzzy classifier and a pixel color changer.
  • the pixel fuzzy classifier determines the fuzzy membership of each pixel in the class of colors to be changed. Once the fuzzy membership of the class of colors to be changed is determined, the pixel fuzzy classifier provides a control signal to the pixel color changer which changes each pixel according to the fuzzy membership of the pixel in the change-class. Pixels close to the change class in color composition are changed according to the fuzzy membership of those pixels in the change-class. As a result, a smooth transition between the change class pixels and the pixels surrounding the change-class can be obtained.
  • FIG. 3 is an exemplary embodiment of the color corrector according to the present invention.
  • the color corrector includes a pixel fuzzy classifier 410 which determines the fuzzy membership of a pixel in the change-class.
  • the exemplary pixel fuzzy classifier includes a neural network.
  • a fuzzy membership value signal cl of a pixel in the change-class is provided to pixel color changer 420.
  • Pixel color changer 420 changes the pixels using either equation (1) or (2) and fuzzy membership value signal.
  • Pixel changer 420 is a non-neural processor.
  • vectors A or V of equations (1) and (2) are separated into three components ⁇ R, ⁇ G and ⁇ B which are separately multiplied by the fuzzy membership value and respectively added to the R, G, and B components of the pixel to be color corrected.
  • Vectors A or V are provided through interface device 405 in color change signal CCS to memory units 430a, 430b, and 430c.
  • Memory units 430 provided the components ⁇ R, ⁇ G and ⁇ B to respective multipliers 435a, 435b, and 435c which provided outputs to respective adders 440a, 440b, and 440c.
  • a second pixel changer 421 can be included for changing a second change-class of colors.
  • the pixel fuzzy classifier 410 provides a second membership value c2 for the pixel providing the pixels fuzzy membership in the second change-class.
  • the second pixel changer 421 changes the pixel using either equation (1) or (2) and fuzzy membership value c2.
  • the components of a second vector A or V in equations (1) and (2), respectively, are multiplied by the fuzzy membership value c2 and then added to the color corrected pixel from the first pixel changer 420.
  • further pixel changers can be provided for additional change-classes.
  • the pixel fuzzy classifier 410 is trained by example to produce fuzzy membership value c for the different pixels in an image.
  • a neural network essentially performs a process called vector mapping where it maps a set of input values to a set of output values as shown in FIG. 4a.
  • FIG. 4a illustrates a conventional feed-forward neural network where the weights, w, scale the inputs to the nodes, or neurons, n. Signals are provided to terminals IN An output signal V i m , not shown produced at node n im at layer m, is determined by equations (3) and (4) below. ##EQU1##
  • w ij m is the strength of the connection from node V j m-1 to V i m
  • w io is the bias of the node
  • N m is the number of nodes in the m th layer.
  • FIG. 4b An exemplary embodiment of the neural network 410a in pixel fuzzy classifier 410 is shown in FIG. 4b in block form.
  • the neural network is provided pixels comprising one or more pixel data values R, B, and G for the three primary colors red, blue, and green.
  • the output signal of the neural network is fuzzy membership values c1 and c2. Fuzzy membership values c1 and c2 range from zero to one where zero represents no membership in the change-class and one corresponds to full membership in the change class.
  • the fuzzy membership value for the pixel applied to the neural network is the value c1. If two change-classes are identified by the neural network, then the membership of the pixel in the first change-class is fuzzy membership value c1 and the membership of the pixel in the second change class is fuzzy membership value c2.
  • fuzzy membership values are shown in FIG. 4b for two different change classes, any number of fuzzy membership values can be produced corresponding to any number of change-classes.
  • an additional output node is provided in the neural network.
  • the strength of the connections between the nodes W i m , as well as the bias of the nodes W io are generated using the flow chart shown in FIG. 6.
  • the neural network "learns" by a process experience arising from exposure to measurements of empirical phenomena which are converted to empirical rules that are embodied in network weights.
  • the pixels for training the pixel fuzzy classifier 410 are provided by a user through interface device 405 in example pixel signal EPS1 or by pixel group classifier 400 in example pixel signal EPS2.
  • Pixel group classifier contains a second neural network for identifying groups of pixels from an image data signal IMAGE -- DATA, provided from the image to be color corrected.
  • the exemplary embodiment of the present invention uses a combination of both supervised and unsupervised learning for training the pixel fuzzy classifier.
  • a Kohonen Neural Network is trained using unsupervised learning to identify groups of pixels to be used to train a feed-forward neural network.
  • the groups of pixels identified by the Kohonen Neural Network are applied to the feed-forward neural network 410a as example pixels and used for supervised training of pixel fuzzy classifier 410.
  • the pixel change-class can be quickly and accurately trained with little or no user involvement.
  • supervised learning there is a training set that consists of a set of input and desired-output vector pairs.
  • the vector pairs are typically used in an iterative process. First, the input vector is applied to the network which produces an output which is then compared to a desired output. Next, the weight values and bias of the node are adjusted to decrease this difference.
  • the iterative process is usually a form of gradient descent where the gradient of the error surface is estimated, and the weight values are adjusted based on the gradient so that the root-mean squared error is minimized.
  • the root-mean squared error is the square root of the square of the difference between the actual result and the desired result.
  • Back propagation and conjugate gradient as are well known in the art are two common examples of iterative processes that rely on gradient information. The conjugate gradient is explained below.
  • Unsupervised learning is normally used to discover unknown patterns in data. There are no desired outputs, only input data. Data is clustered or categorized in these networks. Similar inputs are classified in the same cluster causing the same output node to "fire.”
  • the exemplary embodiment of the present invention uses both unsupervised and supervised learning to identify the pixel change-class.
  • the conjugate gradient search is used to slide the weight vector down the nearest slope. After the gradient technique appears to be at a minimum, simulated annealing is used to try to "bump" the weight vector out of a possible local minimum. If a substantial advantage is achieved by annealing, the algorithm repeats the conjugate gradient search until only trivial improvements are made or until a user set retry limit is reached.
  • the retry limit is chosen to minimize processing time and to achieve a desired improvement in the obtained weights.
  • the weights are improved as the value of the retry limit is increased. As the retry limit is increased, however, the time required to train the neural network increases. Accordingly, the user balances improvement of the weights against time for training. In the exemplary embodiment of the present the retry limit is five.
  • FIG. 6 is a flow chart of the training algorithm for training the feed-forward neural network in fuzzy neural classifier 410.
  • the training method includes three processes: genetic optimization, conjugate gradient search, and simulated annealing. These processes are well known to one of ordinary skill in the art of neural network design as illustrated in Practical Neural Network Recipes in C++, by Timothy Masters, San Diego, Academic Press, Inc. (1993), which is incorporated herein by reference for its teachings on the training of neural networks.
  • example pixels are provided to the neural network 410a by the user or the Kohonen Neural Network.
  • the example pixels include pixels in the change-class as well as pixels in the do-nothing class. To ensure proper training, one third to one half of the total number of example pixels should be pixels in the change-class. If enough pixels in the change-class are not identified, additional change-class pixels are generated from the existing change-class examples by, for example, duplication.
  • Pixels are provided in each class so the feed-forward neural network can be trained to determine the difference between the change-class and the do-nothing class and produce fuzzy membership value c of a pixel in the change-class.
  • the feed-forward neural network can be trained to determine the difference between the change-class and the do-nothing class and produce fuzzy membership value c of a pixel in the change-class.
  • a genetic structure representing the parameters of the optimization problem is chosen.
  • the genetic structure are the weights of the neural network. Genetic optimization is used to initialize the neural network weights to obtain an initial weight vector which is closer to a global minimum than would result from random weight initialization.
  • Initialization is performed by randomly generating a population of weight sets having weight values for the different nodes of the neural network. Ideally the weight values of the weight sets are relatively small random values which are assigned to the different nodes. These initial values represent a starting point in the search for the best set of initial weight values which will produce the correct membership of a pixel in the change-class. Each individual weight set in the population is tested using an objective function to compute a fitness value which is a measure of the optimization of an objective function by the weight set.
  • the objective function used in the exemplary embodiment of the invention is the root-mean squared error.
  • Example pixels are applied to the neural network for each weight set generated to produce a parent fuzzy membership value for each weight set.
  • the "difference of the actual result and the desired result" in the root-mean squared error is the difference between the parent fuzzy membership value and the fuzzy membership value determined according to relation (8), for example, below:
  • PX ⁇ R,G,B ⁇ is the input example pixel having pixel data values R, G, and B
  • PX' ⁇ R',G',B' ⁇ is a pixel in the change-class
  • c is the membership value.
  • an input pixel PX ⁇ R,G,B ⁇ having pixel data values R, G, and B is or is not a member, i.e.--having one hundred percent or zero percent membership, in the change-class defined by pixels PX' ⁇ R',G',B' ⁇ having pixel data values R', G', and B' according to relation (8).
  • weight sets which produce the greatest optimization of the objective function i.e., the smallest difference value
  • parent weight sets are selected and used to generate new weight sets, "children weight sets”.
  • ten weight sets can be randomly generated.
  • Four of the ten weights which generate the lowest root-mean squared error, i.e.--optimize the objective function are selected as parent sets to generate children sets.
  • a child weight set can be generated by, for example, averaging the corresponding weights in each of the four parent sets.
  • the weights can be generated by copying a portion of the weights from one parent weight set to a second parent weight set. For example, if the parent weight sets have 10 weight values, one weight value for each node in the neural network, the first five weight values of the first parent weight set can be combined with the last five weight values of the second parent weight set to produce a child weight set.
  • mutation can be introduced. For example, a random number can be substituted for one of the weights in the child weight set to introduce a mutation in the generation of the child weight set.
  • the conjugate gradient process modifies the weight values of a child weight set generated at step 520 to move the weight values towards the global minimum by applying the example pixels obtained at step 510 to neural network 410a. Then, the conjugate gradient process uses information about the gradient of the error surface to update the weight values in an iterative process. Each step of the iterative process does not undo any error reduction of the previous steps because the direction of the current step is conjugate to the previous steps.
  • a second weight set is applied to the neural network 410a to produce a second pixel fuzzy membership value.
  • a conjugate gradient algorithm as is well known to those skilled in neural network design, for example the Polak-Ribiere algorithm, is applied to determine the search direction to avoid selecting the wrong search direction when the subsequent error value is generated. The search direction is selected to minimize the root-mean square error. The root-mean square error is then determined for this value. This process is repeated until the root-mean error produced for a corresponding weight set does not decrease.
  • a first weight set is used to produce error value 610
  • a second weight set is used to produce error value 620
  • a third weight set is used to produce error value 630
  • a fourth weight set is used to produce error value 640.
  • Error value 640 has not decreased from the previous error value 630. Accordingly, a minimum exists between error value 620 and 640.
  • an error value 650 is generated between error values 630 and 640 using a fifth weight set. Since error value 650 is less than error value 640 but greater than error value 630, error value 640 is discarded. The minimum is between error values 620 and 650. This process is repeated to obtain error value 660 which is a minimum corresponding to a sixth weight set.
  • simulated annealing is performed at step 540 on the weight set which produced the minimum error value, for example, the sixth weight set. Simulated annealing is performed by (1) randomly perturbing the weight values in the neural network, and (2) tracking the lowest error value produced by perturbing the weight values. The weight values are perturbed using a random number generator having a relatively low standard deviation.
  • step 550 it is determined in the simulated annealing produced improved weight values. If the weight values are improved, it is then determined at step 560 if the improvement is trivial. If the improvement is not trivial, then step 530 is repeated to perform the conjugate gradient for the weight values generated at step 540. After many tries, the set of weight values that produce the best function value, i.e.--the weight values which minimize the root-mean squared error, are designated to be the center about which perturbation will take place for the next desired state. The standard deviation of the random number generator for the next desired state is then reduced, and simulated annealing is performed again. If the improvement is trivial as determined at step 560 or if the weight values are not improved at step 550, step 570 is performed.
  • step 570 it is determined whether a specified number of trials has been reached, for example, five. If so, the training method is terminated. If, however, the trial limit has not been reached, then at step 580 the weights are set at the state prior to the start of the conjugate gradient step 530. Then step 530 is repeated.
  • the trial limit can be set by the user. The weights from the above process which produce the lowest root-mean squared error are selected as the weights for the neural network to classify the fuzzy membership c of pixels in the change-class.
  • FIG. 7 illustrates how the training method in FIG. 6 can proceed.
  • weight values 710a-710e are identified and weight value 710c at the lowest minimum is selected as the initial weight value for step 530.
  • local minimum 720 is identified.
  • simulated annealing is performed at step 540 which breaks out of the local minimum to weight value 730.
  • conjugate gradient step 530 is then performed again to find the next local minimum 740. In this case the local minimum and the global minimum are the same.
  • FIG. 8 is a three-dimensional plot of the color composition of an image in an RGB cube. Only the luminance component of the image is shown in the figures. The pixels in the change-class are identified by the darker dots 800.
  • Example pixels in the class or classes of colors to be changed, the change-class, are provided to the fuzzy neural classifier 410 by pixel group classifier 400 or interface device 405.
  • the neural classifier 410 is also provided the direction in which the pixels should be changed using interface device 405.
  • Table 1 shows a list of example pixels and the vector to change the pixels.
  • ⁇ rll,gll,bll ⁇ is an example pixel including red r, green g, and blue b components
  • ⁇ rvl,gvl,bvl ⁇ is a vector specifying the amount each individual component is to be changed.
  • pixel the ⁇ rll,gll,bll ⁇ is to be change to ⁇ (rll+rv1), (gll+gv1),(bll+bv1) ⁇ .
  • examples of pixels not belonging to the change-class are also provided to the fuzzy neural classifier from pixel group classifier 400 or interface device 405.
  • the pixels not in the change class are called the do-nothing class.
  • One third to one half of the example pixels provided to the neural network for training should be in the change-class. If there are not enough pixels in the change class to satisfy the above quantity, the system can include a system for duplicating the pixels in the change-class so that an adequate number of pixels can be provided to the neural network for training.
  • the do-nothing class is represented as a arbitrary sampling of all colors in an RGB cube which are sampled linearly in each dimension.
  • the change-class examples and the do-nothing class examples are projected onto an RG plane as illustrated in FIG. 9.
  • the do-nothing examples are represented by large open dots 805, the change-class examples are represented by large black dots 800, and the image pixels as small dots 810.
  • a grid of do-nothing examples is used and examples of specifically what not to include are not provided to the pixel fuzzy classifier.
  • the example pixels in FIG. 9 have been provided to the fuzzy pixel classifier 410 through interface device 405.
  • Pixels in the change-class near the arbitrarily chosen pixels in the do-nothing-class may be incorrectly included in the do-nothing class. This problem can be avoided by creating a preponderance of pixels in the change-class, so that the root-mean squared error favors the chosen pixels in the change-class.
  • the distinction between the change-class and the do-nothing class is improved by finding clusters of points in the original image and using those points as examples of the do-nothing class.
  • the new do-nothing examples 815 shown in FIG. 10 are used to improve the fuzzy membership of the pixels in the do-nothing class and the change-class.
  • the do-nothing and the change-class pixels from FIG. 10 are used at step 520, shown in FIG. 6, to initialize the weights of the neural network.
  • FIGS. 13 illustrates another example of changing pixels in a change-class using the exemplary embodiment of the present invention.
  • FIG. 13 is a plot showing the projection of the location of each pixel in an image onto a two-dimensional plane.
  • the examples in the training set for each red-green pair is repeated to cover the blue dimension.
  • the training set in the change-class 1300 is shown as solid dots. For the point shown at (50,50) the training set actually included the points (50,50,50), (50,50,150) and (50,50,250).
  • the feed-forward neural network has a 3-4-1 architecture having three inputs for receiving the pixel data values R, G, and B, four nodes, and one fuzzy membership value c1. This the same neural network shown in FIG. 4b without the additional node for fuzzy membership value c2. Only one fuzzy membership value c1 is used since there is only one change class.
  • the fuzzy neural classifier was trained and executed on an image that sampled the Red-Green-Blue cube at 10 equally spaced points along each dimension, yielding 1000 remapped points. The projection of these points onto the Red-Green plane is shown in FIG. 14 with the points connected in a grid to emphasize the remapping. The deviation of an intersection from a rectangular grid point is the extent of color correction at that point.
  • the fuzzy neural classifier 410 changed pixels in or near the change-class in the proper direction. For example, the pixels in the upper left-hand corner of the Red-Green plane were left unchanged.
  • the fuzzy classifier network 410 computed the fuzzy membership value for the pixels, leaving the actual remapping to the pixel changer.
  • a Kohonen Neural Network using unsupervised training can be used to find clusters in the color space of the image to provide do-nothing examples to train the feed-forward neural network 410a.
  • the image data values R, G, and B of the clusters are provided to the feed-forward neural network 410a as example pixels.
  • the clusters generated by the Kohonen Neural Network also include a cluster corresponding to the change-class of the neural network.
  • the cluster corresponding to the change-class is also provided to the neural as a do-nothing example. This may introduce a minor error in weight values produced when training the neural network. This error can be ignored when the number of clusters identified by the neural network is much greater than one, for example, fifty.
  • the cluster corresponding to the change-class can be identified by comparing the clusters to the change-class pixels and not providing the identified cluster as a do-nothing example to the feed-forward neural network for training.
  • the Kohonen Neural Network shown in FIG. 11 is an example of an unsupervised learning network which is included in the pixel group classifier 400.
  • a Kohonen Neural Network is capable of finding patterns of m multi-dimensional data. Accordingly, clusters of pixels in the RGB color space of an image can be identified. The image under consideration is the training set for the Kohonen Neural Network. Once the network is trained, the weights to the output nodes are the locations of the clusters.
  • the Kohonen Neural Network operates on the principal of competitive learning, where output neurons "compete” to win. During training, only one, or sometimes a few neurons "win” the right to have their weights updated.
  • the exemplary Kohonen Neural Network is provided image pixels comprising one or more pixel data values R, B, and G for the three primary colors red, blue, and green from the image.
  • the pixel data values R, G, and B are first normalized in normalizing unit 1100.
  • Next pixel data values R, G, and B are multiplied by a weight w ij to produce an output data value h i .
  • the output data value, h i of each unit in the Kohonen Neural Network is determined by equation (5) below:
  • w ij is the weight from the jth input to the ith output node and I j is the jth input which is one of the pixel data values R, G, and B.
  • the winner is the node with the greatest data value output h i .
  • the winning node has its weight updated according to equation (6) below:
  • ⁇ w ij is the amount the weight from the jth input to the ith output node, the node with the greatest output, is changed and ⁇ is a fixed constant used to update the weights. ⁇ is the learning rate of the network.
  • the result of equation (6) is used to move the weights of the winning data value output h i closer to the input pixel data values R, G, and B by adding the value of ⁇ w ij to the value of the weight of the jth input to the ith output node.
  • the weights themselves represent the locations of the clusters.
  • Pixel group classifier 400 is shown in FIG. 3 as part of the non-linear color corrector, pixel group classifier 400 need not be included in the color correction system. Pixel group classifier 400 can operate independently of the non-linear color corrector.
  • pixel group classifier 400 provides the weights from the Kohonen Neural Network to the neural network classifier, which uses the weights as pixel data values to train feed-forward neural network 410a.
  • FIG. 12 illustrates example pixels to be provided to the fuzzy neural classifier for training.
  • the change-class examples are illustrated by dark spots 800, the Kohonen Neural Network located clusters are illustrated by squares 820, and the pixels by smaller dots 810.
  • the number of examples in the do-nothing class where reduced by using the Kohonen Neural Network to identify the cluster of pixels.
  • pixel fuzzy classifier 410 and pixel color changer 420 can be combined.
  • the neural network classifies pixels and remaps the pixels using a neural network.
  • FIG. 15 is another exemplary embodiment of the present invention.
  • FIG. 15 is a color corrector system for correcting the color of a video image provide by a video tape player (VTP) 1500.
  • the video from VTP 1500 is, for example, a film that has been transferred to video.
  • the colors which look suitable when projected from the film onto a screen are not necessarily suitable for display on a video screen from a video tape. Accordingly, it is necessary to color correct scenes in the film which have been transferred to video tape so that the picture produced from the film is suitable for display on a video screen.
  • VTP 1500 provides a video signal from a video tape to display control unit 1510.
  • Display control unit 1510 displays an image on displays 1520 and 1530 corresponding to the video signal.
  • the displayed image would be, for example, one scene.
  • a scene may last only for a few moments or for a longer duration.
  • the user first color corrects a frame of the scene as displayed. Once that frame has been corrected to the user's satisfaction, the entire scene is color corrected.
  • the user after reviewing the displayed image frame identifies pixels in the image to be changed using keyboard 1540 and mouse 1550.
  • the user would specify the color the change-class pixels are to be changed. For example, the user could identify light red pixels to be changed to dark red pixels.
  • the system would then determine the appropriate values for ⁇ R, ⁇ G, and ⁇ B used by the color changer. Then these values are provided to the color corrector 1560 which is the color corrector shown in FIG. 3.
  • the user selects several pixels in the change-class by pointing with the mouse and marking the pixels to be changed.
  • the user may identify pixels in the do-nothing class.
  • the change-class examples and the do-nothing examples are then provided to the color corrector which uses the examples to train the neural network.
  • the user can identify pixels in the do-nothing class or the Kohonen Neural Network can produce the do-nothing examples.
  • the color corrector includes a Kohonen network for generating examples of pixels in the do-nothing class.
  • the display control unit 1510 provides the pixels from the scene to be color corrected to color corrector 1560.
  • Color corrector 1560 calculates the fuzzy membership value c for each pixel in the scene and changes the pixel according to the pixels membership in the change-class and the color to which the change-class is to be changed.
  • the color corrected pixels are provided to the display control unit 1510 and displayed on display 1530.
  • the user can examine the difference between the corrected image displayed on display 1530 and the original image displayed on display 1520. If the change is not acceptable, the user can change the change-class by identifying new pixels in the original image or change the color to which the change-class is to be changed. If a new change-class is identified, display control unit 1510 provides the new change-class to color corrector 1560 to retrain the neural network. Then the pixels in the scene are color corrected. Alternatively, if the color is only changed, the pixels in the scene are provided to the color corrector which changes the pixels in the original change class according to the new color specified by the user.
  • the entire scene is color corrected by providing the image signal to display control unit 1510 which provides the image signal to color corrector 1560.
  • Color corrector classifies the pixels in the image and color corrects the pixels in the image using the membership value of the pixels in the change-class.
  • the color corrected pixels are provided to display control unit 1510 which provides the color corrected scenes to video tape recorder (VTR) 1570 which records the color corrected video.
  • VTR video tape recorder

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)
  • Processing Of Color Television Signals (AREA)
  • Facsimile Image Signal Circuits (AREA)
  • Color Image Communication Systems (AREA)
US08/367,798 1994-12-30 1994-12-30 Non-linear color corrector having a neural network and using fuzzy membership values to correct color and a method thereof Expired - Fee Related US5548697A (en)

Priority Applications (4)

Application Number Priority Date Filing Date Title
US08/367,798 US5548697A (en) 1994-12-30 1994-12-30 Non-linear color corrector having a neural network and using fuzzy membership values to correct color and a method thereof
EP95120188A EP0720390B1 (en) 1994-12-30 1995-12-20 Non-linear color corrector and method thereof
DE69517401T DE69517401T2 (de) 1994-12-30 1995-12-20 Nichtlineare Farbkorrekturvorrichtung und geeignetes Verfahren
JP7341926A JPH08256349A (ja) 1994-12-30 1995-12-27 非線形色補正装置および方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
US08/367,798 US5548697A (en) 1994-12-30 1994-12-30 Non-linear color corrector having a neural network and using fuzzy membership values to correct color and a method thereof

Publications (1)

Publication Number Publication Date
US5548697A true US5548697A (en) 1996-08-20

Family

ID=23448654

Family Applications (1)

Application Number Title Priority Date Filing Date
US08/367,798 Expired - Fee Related US5548697A (en) 1994-12-30 1994-12-30 Non-linear color corrector having a neural network and using fuzzy membership values to correct color and a method thereof

Country Status (4)

Country Link
US (1) US5548697A (enrdf_load_stackoverflow)
EP (1) EP0720390B1 (enrdf_load_stackoverflow)
JP (1) JPH08256349A (enrdf_load_stackoverflow)
DE (1) DE69517401T2 (enrdf_load_stackoverflow)

Cited By (27)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5652831A (en) * 1996-02-01 1997-07-29 Industrial Technology Reasearch Institute Variable point interpolation apparatus and method with scalable architecture for color correction
US5778156A (en) * 1996-05-08 1998-07-07 Xerox Corporation Method and system for implementing fuzzy image processing of image data
US6023663A (en) * 1996-04-08 2000-02-08 Samsung Electronics Co., Ltd. Method and apparatus for inspecting a solder joint using a correlation neural network
US6058207A (en) * 1995-05-03 2000-05-02 Agfa Corporation Selective color correction applied to plurality of local color gamuts
US6185336B1 (en) * 1998-09-23 2001-02-06 Xerox Corporation Method and system for classifying a halftone pixel based on noise injected halftone frequency estimation
US6233365B1 (en) * 1996-05-27 2001-05-15 Sharp Kabushiki Kaisha Image-processing method
US20020012073A1 (en) * 2000-07-28 2002-01-31 Fujitsu Hitachi Plasma Display Limited Color reproduction correction circuit for color representation and a color display
US20020102018A1 (en) * 1999-08-17 2002-08-01 Siming Lin System and method for color characterization using fuzzy pixel classification with application in color matching and color match location
US20020118883A1 (en) * 2001-02-24 2002-08-29 Neema Bhatt Classifier-based enhancement of digital images
US20020184167A1 (en) * 2001-06-05 2002-12-05 Mcclanahan Craig J. System and method for converting a color formula using an artificial intelligence based conversion model
US20020184168A1 (en) * 2001-06-05 2002-12-05 Mcclanahan Craig J. System and method for determining acceptability of proposed color solution using an artificial intelligence based tolerance model
US20030068080A1 (en) * 2001-08-08 2003-04-10 Chung-Yen Lu Color image processing apparatus and method for color enhancement
US20030147558A1 (en) * 2002-02-07 2003-08-07 Loui Alexander C. Method for image region classification using unsupervised and supervised learning
US20030185457A1 (en) * 2002-03-29 2003-10-02 Campbell Richard John Methods and systems for adjusting digital image characteristics
US6803956B1 (en) * 2000-06-12 2004-10-12 Pulnix America, Inc. Color recognition camera
US6999615B2 (en) 2001-06-05 2006-02-14 Basf Corporation Color management and solution distribution system and method
US20060089924A1 (en) * 2000-09-25 2006-04-27 Bhavani Raskutti Document categorisation system
US20060153439A1 (en) * 2005-01-11 2006-07-13 Omron Corporation PC board inspecting apparatus, inspection logic setting method, and inspection logic setting apparatus
US7233692B2 (en) 2002-11-14 2007-06-19 Lockheed Martin Corporation Method and computer program product for identifying output classes with multi-modal dispersion in feature space and incorporating multi-modal structure into a pattern recognition system
US7405741B1 (en) * 2004-08-31 2008-07-29 Pixelworks, Inc. Fuzzy logic based LCD overdrive control method
US7427993B1 (en) * 2004-08-31 2008-09-23 Pixelworks, Inc. Motion adaptive pixel boost with data compression and decompression
US7443370B1 (en) 2004-08-31 2008-10-28 Pixelworks, Inc. YUV compression for boost
US7570809B1 (en) * 2004-07-03 2009-08-04 Hrl Laboratories, Llc Method for automatic color balancing in digital images
US9123087B2 (en) 2013-03-25 2015-09-01 Xerox Corporation Systems and methods for segmenting an image
US20170185871A1 (en) * 2015-12-29 2017-06-29 Qiang Zhang Method and apparatus of neural network based image signal processor
US11003989B2 (en) * 2017-04-27 2021-05-11 Futurewei Technologies, Inc. Non-convex optimization by gradient-accelerated simulated annealing
US20230316094A1 (en) * 2022-03-29 2023-10-05 D-Wave Systems Inc. Systems and methods for heuristic algorithms with variable effort parameters

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7144474B1 (en) 1992-08-17 2006-12-05 Weyerhaeuser Co. Method of binding particles to binder treated fibers
JP2000311243A (ja) * 1999-04-28 2000-11-07 Sony Corp 画像色補正方法及び装置
JP4491988B2 (ja) * 2001-04-09 2010-06-30 パナソニック株式会社 カラー画像処理方式及びカラー画像処理装置
CN109767399B (zh) * 2019-01-14 2023-04-07 哈尔滨工程大学 一种基于无监督色彩校正的水下图像增强方法
DE102019209282A1 (de) * 2019-06-26 2020-12-31 Robert Bosch Gmbh Verfahren zum Erhalten einer Infrarotdarstellung einer Bildsituation und Recheneinheit, System und Computerprogramm zu dessen Durchführung
CN116878335A (zh) * 2023-08-08 2023-10-13 青岛九维华盾科技研究院有限公司 一种基于分布式无线网络的自适应伪装模块激发控制系统

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5162899A (en) * 1989-03-15 1992-11-10 Matsushita Electric Industrial Co., Ltd. Color data correction apparatus ultilizing neural network
US5185850A (en) * 1990-05-22 1993-02-09 Toyo Ink Mfg. Co., Ltd. Color transformation method and apparatus for transforming physical to psychological attribute using a neural network
US5255344A (en) * 1988-05-20 1993-10-19 Matsushita Electric Industrial Co., Ltd. Inference rule determining method and inference device
US5285297A (en) * 1991-06-25 1994-02-08 Scitex Corporation Ltd. Apparatus and method for color calibration
US5289295A (en) * 1991-07-04 1994-02-22 Matsushita Electric Industrial Co., Ltd. Color adjustment apparatus
US5339365A (en) * 1989-08-31 1994-08-16 Canon Kabushiki Kaisha Image processing apparatus capable of using fuzzy logic
US5386496A (en) * 1991-02-15 1995-01-31 Toyo Ink Mfg. Co., Ltd. Method and device for nonlinear transformation of colour information by neural network
US5425108A (en) * 1992-09-04 1995-06-13 Industrial Technology Research Institute Mobile type of automatic identification system for a car plate
US5434927A (en) * 1993-12-08 1995-07-18 Minnesota Mining And Manufacturing Company Method and apparatus for machine vision classification and tracking
US5444796A (en) * 1993-10-18 1995-08-22 Bayer Corporation Method for unsupervised neural network classification with back propagation

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH05276371A (ja) * 1992-03-27 1993-10-22 Sanyo Electric Co Ltd カラー画像処理方法

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5255344A (en) * 1988-05-20 1993-10-19 Matsushita Electric Industrial Co., Ltd. Inference rule determining method and inference device
US5162899A (en) * 1989-03-15 1992-11-10 Matsushita Electric Industrial Co., Ltd. Color data correction apparatus ultilizing neural network
US5339365A (en) * 1989-08-31 1994-08-16 Canon Kabushiki Kaisha Image processing apparatus capable of using fuzzy logic
US5185850A (en) * 1990-05-22 1993-02-09 Toyo Ink Mfg. Co., Ltd. Color transformation method and apparatus for transforming physical to psychological attribute using a neural network
US5386496A (en) * 1991-02-15 1995-01-31 Toyo Ink Mfg. Co., Ltd. Method and device for nonlinear transformation of colour information by neural network
US5285297A (en) * 1991-06-25 1994-02-08 Scitex Corporation Ltd. Apparatus and method for color calibration
US5289295A (en) * 1991-07-04 1994-02-22 Matsushita Electric Industrial Co., Ltd. Color adjustment apparatus
US5425108A (en) * 1992-09-04 1995-06-13 Industrial Technology Research Institute Mobile type of automatic identification system for a car plate
US5444796A (en) * 1993-10-18 1995-08-22 Bayer Corporation Method for unsupervised neural network classification with back propagation
US5434927A (en) * 1993-12-08 1995-07-18 Minnesota Mining And Manufacturing Company Method and apparatus for machine vision classification and tracking

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
P. Wasserman, "Advanced Methods in Neural Computing", published by New York Van Nostrand Reinhold (1993).
P. Wasserman, Advanced Methods in Neural Computing , published by New York Van Nostrand Reinhold (1993). *
T. Masters, "Practical Neural Network Recipes in C++", San Diego, Academic Press, Inc. (1993), Chapters 6, 8, 17 & 18, and pp. 21, 22, 409 & 411.
T. Masters, Practical Neural Network Recipes in C , San Diego, Academic Press, Inc. (1993), Chapters 6, 8, 17 & 18, and pp. 21, 22, 409 & 411. *

Cited By (43)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6058207A (en) * 1995-05-03 2000-05-02 Agfa Corporation Selective color correction applied to plurality of local color gamuts
US5652831A (en) * 1996-02-01 1997-07-29 Industrial Technology Reasearch Institute Variable point interpolation apparatus and method with scalable architecture for color correction
US6023663A (en) * 1996-04-08 2000-02-08 Samsung Electronics Co., Ltd. Method and apparatus for inspecting a solder joint using a correlation neural network
US5778156A (en) * 1996-05-08 1998-07-07 Xerox Corporation Method and system for implementing fuzzy image processing of image data
US6233365B1 (en) * 1996-05-27 2001-05-15 Sharp Kabushiki Kaisha Image-processing method
US6185336B1 (en) * 1998-09-23 2001-02-06 Xerox Corporation Method and system for classifying a halftone pixel based on noise injected halftone frequency estimation
US20020102018A1 (en) * 1999-08-17 2002-08-01 Siming Lin System and method for color characterization using fuzzy pixel classification with application in color matching and color match location
US7046842B2 (en) * 1999-08-17 2006-05-16 National Instruments Corporation System and method for color characterization using fuzzy pixel classification with application in color matching and color match location
US20040228526A9 (en) * 1999-08-17 2004-11-18 Siming Lin System and method for color characterization using fuzzy pixel classification with application in color matching and color match location
US6803956B1 (en) * 2000-06-12 2004-10-12 Pulnix America, Inc. Color recognition camera
US20020012073A1 (en) * 2000-07-28 2002-01-31 Fujitsu Hitachi Plasma Display Limited Color reproduction correction circuit for color representation and a color display
US6940559B2 (en) 2000-07-28 2005-09-06 Fujitsu Hitachi Plasma Display Limited Color reproduction correction circuit for color representation and a color display
US6753930B2 (en) * 2000-07-28 2004-06-22 Fujitsu Hitachi Plasma Display Limited Color reproduction correction circuit for color representation and a color display
US20040218101A1 (en) * 2000-07-28 2004-11-04 Fujitsu Hitachi Plasma Display Limited Color reproduction correction circuit for color representation and a color display
US7971150B2 (en) 2000-09-25 2011-06-28 Telstra New Wave Pty Ltd. Document categorisation system
US20060089924A1 (en) * 2000-09-25 2006-04-27 Bhavani Raskutti Document categorisation system
US20020118883A1 (en) * 2001-02-24 2002-08-29 Neema Bhatt Classifier-based enhancement of digital images
US20050240543A1 (en) * 2001-06-05 2005-10-27 Basf Corporation Method for determining acceptability of proposed color solution using an artificial intelligence model
US20020184168A1 (en) * 2001-06-05 2002-12-05 Mcclanahan Craig J. System and method for determining acceptability of proposed color solution using an artificial intelligence based tolerance model
US20020184167A1 (en) * 2001-06-05 2002-12-05 Mcclanahan Craig J. System and method for converting a color formula using an artificial intelligence based conversion model
US6993512B2 (en) 2001-06-05 2006-01-31 Basf Corporation System and method for converting a color formula using an artificial intelligence based conversion model
US6999615B2 (en) 2001-06-05 2006-02-14 Basf Corporation Color management and solution distribution system and method
US7536231B2 (en) * 2001-06-05 2009-05-19 Basf Corporation Method for determining acceptability of proposed color solution using an artificial intelligence model
US6904165B2 (en) * 2001-08-08 2005-06-07 Silicon Integrated Systems Corporation Color image processing apparatus and method for color enhancement
US20030068080A1 (en) * 2001-08-08 2003-04-10 Chung-Yen Lu Color image processing apparatus and method for color enhancement
US7039239B2 (en) * 2002-02-07 2006-05-02 Eastman Kodak Company Method for image region classification using unsupervised and supervised learning
US20030147558A1 (en) * 2002-02-07 2003-08-07 Loui Alexander C. Method for image region classification using unsupervised and supervised learning
US7221807B2 (en) * 2002-03-29 2007-05-22 Sharp Laboratories Of America, Inc. Methods and systems for digital image characteristic adjustment using a neural network
US20030185457A1 (en) * 2002-03-29 2003-10-02 Campbell Richard John Methods and systems for adjusting digital image characteristics
US7233692B2 (en) 2002-11-14 2007-06-19 Lockheed Martin Corporation Method and computer program product for identifying output classes with multi-modal dispersion in feature space and incorporating multi-modal structure into a pattern recognition system
US7570809B1 (en) * 2004-07-03 2009-08-04 Hrl Laboratories, Llc Method for automatic color balancing in digital images
US7427993B1 (en) * 2004-08-31 2008-09-23 Pixelworks, Inc. Motion adaptive pixel boost with data compression and decompression
US7443370B1 (en) 2004-08-31 2008-10-28 Pixelworks, Inc. YUV compression for boost
US7405741B1 (en) * 2004-08-31 2008-07-29 Pixelworks, Inc. Fuzzy logic based LCD overdrive control method
US8004544B1 (en) 2004-08-31 2011-08-23 Pixelworks, Inc. Fuzzy logic based LCD overdrive control method
US20060153439A1 (en) * 2005-01-11 2006-07-13 Omron Corporation PC board inspecting apparatus, inspection logic setting method, and inspection logic setting apparatus
US9123087B2 (en) 2013-03-25 2015-09-01 Xerox Corporation Systems and methods for segmenting an image
US20170185871A1 (en) * 2015-12-29 2017-06-29 Qiang Zhang Method and apparatus of neural network based image signal processor
KR20170078516A (ko) * 2015-12-29 2017-07-07 삼성전자주식회사 신경망 기반 영상 신호 처리를 수행하는 방법 및 장치
US10460231B2 (en) * 2015-12-29 2019-10-29 Samsung Electronics Co., Ltd. Method and apparatus of neural network based image signal processor
US11003989B2 (en) * 2017-04-27 2021-05-11 Futurewei Technologies, Inc. Non-convex optimization by gradient-accelerated simulated annealing
US20230316094A1 (en) * 2022-03-29 2023-10-05 D-Wave Systems Inc. Systems and methods for heuristic algorithms with variable effort parameters
US12254418B2 (en) * 2022-03-29 2025-03-18 D-Wave Systems Inc. Systems and methods for heuristic algorithms with variable effort parameters

Also Published As

Publication number Publication date
DE69517401T2 (de) 2001-01-25
DE69517401D1 (de) 2000-07-13
JPH08256349A (ja) 1996-10-01
EP0720390A3 (enrdf_load_stackoverflow) 1996-07-24
EP0720390B1 (en) 2000-06-07
EP0720390A2 (en) 1996-07-03

Similar Documents

Publication Publication Date Title
US5548697A (en) Non-linear color corrector having a neural network and using fuzzy membership values to correct color and a method thereof
US7180629B1 (en) Methods and apparatus for color device characterization
US5130935A (en) Color image processing apparatus for extracting image data having predetermined color information from among inputted image data and for correcting inputted image data in response to the extracted image data
US5012333A (en) Interactive dynamic range adjustment system for printing digital images
US6088475A (en) Method and apparatus for forming and correcting color image
CN112785534A (zh) 一种动态场景下去鬼影多曝光图像融合方法
Zamir et al. Vision models for wide color gamut imaging in cinema
JPH04213274A (ja) 自動色補正による電子式画像処理方法及び装置
US6757010B2 (en) Method and system for calibrating color correction instructions between color correction devices
JPH04144361A (ja) セットアップパラメータ決定特性を修正する方法及び自動セットアップ装置
JP2978615B2 (ja) カラーバランス調整装置及び方法
JP3393679B2 (ja) 色画像処理装置
EP0428790B1 (en) Image reproducing method
DeMarsh et al. Color science for imaging systems
EP0940773A2 (en) Image processing method and image processing apparatus
Wang Stoclustic (stochastic clustered) halftone screen design
JP3093220B2 (ja) カラー画像処理方法
JPH07220066A (ja) 画像処理装置
JP3411847B2 (ja) 入力機器用色再現方法及び装置、並びにこの方法のプログラムを記録した記録媒体
JP2958325B2 (ja) 画像処理方法
JPH05207501A (ja) 光源推定装置
CN117979179A (zh) 一种色彩校正方法及系统
CN118612557A (zh) 图像信号处理的训练方法、系统及存储介质
JPH08262692A (ja) 色分解方法および装置
JPH06189125A (ja) ニューラルネットワークによる画像入力装置の色補正回路

Legal Events

Date Code Title Description
AS Assignment

Owner name: PANASONIC TECHNOLOGIES, INC., NEW JERSEY

Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:ZORTEA, ANTHONY E.;REEL/FRAME:007379/0793

Effective date: 19950314

FPAY Fee payment

Year of fee payment: 4

AS Assignment

Owner name: MATSUSHITA ELECTRIC CORPORATION OF AMERICA, NEW JE

Free format text: MERGER;ASSIGNOR:PANASONIC TECHNOLOGIES, INC.;REEL/FRAME:012211/0907

Effective date: 20010928

REMI Maintenance fee reminder mailed
LAPS Lapse for failure to pay maintenance fees
FP Lapsed due to failure to pay maintenance fee

Effective date: 20040820

STCH Information on status: patent discontinuation

Free format text: PATENT EXPIRED DUE TO NONPAYMENT OF MAINTENANCE FEES UNDER 37 CFR 1.362